PyCDA icon indicating copy to clipboard operation
PyCDA copied to clipboard

Published 20 hours ago • verified publisher icon lianqing11

Pre-training

Open Xcrid opened this issue 4 years ago • 2 comments

Hi,

When you pre-train your model with the source dataset, do you also pre-train the aux_output by using the aux_loss or do you only use the main loss ?

Xcrid avatar May 26 '20 14:05 Xcrid

Hi,

I use the aux loss and the loss weight for aux loss is 0.4. Actually I try both, with or without aux loss get similar results. The gap between them is smaller than the randomness in the target domain. You can try the source only with aux loss by modifying train_adabn.py (just delete the target domain part). If you find that I do not reply to you in time, you can send an email to [email protected] since GitHub does not has the message notification.

Best regards, Qing LIAN

lianqing11 avatar May 27 '20 05:05 lianqing11

Thanks, I'll try that

Xcrid avatar May 27 '20 08:05 Xcrid